Can tool schemas become a small model's routing questions directly?

Jev × WebMCP generates questions from tool descriptions and parameter constraints exposed by a web page: tool names and enumerations become multiple-choice questions, booleans become yes-or-no judgments, and free-text values are selected from excerpts of the user's own words. Missing parameters can be left for the user to supply. The extension distinguishes read-only, state-changing, and high-consequence actions. High model confidence does not remove the confirmation required for the latter two categories. This design narrows the scope for freely generating parameters, but tool descriptions still come from the page and should not therefore be treated as trusted instructions. The tradeoff highlighted by the editors is that candidate constraints can reduce formatting errors without guaranteeing truthful tool descriptions, correct parameter semantics, or sufficient permissions. Screening descriptions can help, but permission checks, parameter validation, and user confirmation before execution should still be enforced by the program. As the tool catalog grows, it is also worth considering whether narrowing the candidates first might exclude the correct tool, and whether “no suitable tool” is allowed as an ordinary outcome. Discussion questions 1. As the tool catalog grows, which long-tail tools might a retrieve-then-select approach miss? 2. Should free-text parameters be copied from the user's words, produced by a generative model, or supplied by the user when missing? 3. Which actions must retain independent confirmation even when confidence is high? Source: Jev × WebMCP extension original README https://github.com/sdras/jev-webmcp-extension/blob/main/README.md Editorial note: The official S1 Hub editorial account used AI assistance to compile public sources. This was not submitted by the original author, and this site has not conducted independent hands-on testing. Additional evidence and differing views are welcome.